System Architecture

The Three-Tier Resolution Pipeline & Living Shards

When an inquiry enters MBHub, the client executes a graduated 3-layer resolution pipeline designed to resolve answers as close to your machine as possible. A question solved once is sealed, preserved in the decentralized commons, and served forever at zero marginal cost.

Query Resolution Flowchart

Input & Normalization

Unicode NFKC lowercasing, punctuation stripping, and 64-bit SimHash coordinate generation.

Pre-Flight DLP Gate

Microsecond regex & Luhn checks guarantee zero credential or token leakage.

L1: Local Shard Lookup

Embedded SQLite search via Hamming distance ($H \le 3$, ≥ 85% match) in 0 to 5 ms.

L2: P2P Swarm Gossip

Noise-encrypted GossipSub retrieval from worldwide peers with 600 ms timeout and jitter.

L3: BYOK Live Fallback

Novel queries trigger direct TLS 1.3 streaming from your provider and broadcast to swarm.

1. The Graduated Resolution Tiers

Layer 1

Local Shard (Embedded SQLite)

0 – 5 ms Latency

Your node maintains an embedded, high-concurrency SQLite database operating in Write-Ahead Logging (WAL) mode. When you submit an atomic query, MBHub computes its 64-bit SimHash fingerprint and scans the local cache.

Latency: Instantaneous (< 5 ms), zero network I/O.
Matching: Hamming distance metric over 64-bit hyperplanes.
Offline Capability: 100% functional on airplanes, off-grid, or during outages.
Cost: Exactly $0.00 and 0 tokens.

Layer 2

P2P Swarm (Decentralized Mesh)

≤ 600 ms Latency

If the local shard suffers a cache miss, MBHub queries neighboring peers over a decentralized libp2p GossipSub mesh. All peer-to-peer traffic travels over Noise-protocol encrypted TCP streams with ephemeral peer IDs.

Latency: Bound by a strict 600 ms timeout ceiling.
De-correlation Jitter: 50 to 300 ms randomized timing offsets protect query origin.
Wire Security: Strict 64 KB packet ceiling and canonical BLAKE3 verification.
Cost: $0.00 — mutual peer resource sharing.

Layer 3

BYOK Live Frontier Inference

BYOK Fallback

When an inquiry has never been asked by any node in the global collective, MBHub falls back to your configured commercial AI provider (OpenAI, Anthropic, DeepSeek, Gemini, Groq).

Direct TLS 1.3: Connection connects straight from your machine to provider API endpoints.
Zero Brokerage: No middleman proxies, no markup, no tokens logged by MBHub.
Collective Sealing: Once received, the answer is cryptographically sealed with BLAKE3 and shared to the swarm so nobody ever pays for it again.

2. SimHash & Semantic Fingerprinting

Natural language varies: "How do I reverse a list in Python?" and "Python reverse list syntax" express identical intent. Traditional cryptographic hashes (SHA-256) change completely with a single byte alteration. MBHub utilizes Locality-Sensitive Hashing (SimHash):

The SimHash Coordinate Algorithm

The query string is normalized via Unicode NFKC, converted to lowercase, and stripped of extraneous punctuation.
The text is tokenized into word and character n-grams, each mapped to a 64-bit hash.
A 64-dimensional accumulator vector is updated: each bit position increments by token weight for 1, or decrements for 0.
The final 64-bit SimHash coordinate is produced: bit $i$ is 1 if accumulator[i] > 0, else 0.

Hamming Distance Metric: D_H(A, B) = popcount(A ⊕ B). A Hamming distance of ≤ 3 indicates semantic similarity ≥ 85%. Inquiries within this threshold resolve instantly from L1 cache without network calls.

3. Living Shards: Storage Organism & Memory Modes

In MBHub, the embedded SQLite database is an active organism that continually curates knowledge according to the operator's preferences:

Query Locality Mode (Default)

Your node listens to GossipSub traffic, magnetizing incoming answers that share semantic proximity with topics you frequently research. Over time, your local shard organically crystallizes into a dense, bespoke domain library (e.g. Rust, cryptography, systems programming).

Ideal for developers & researchers

Blind Swarm Mode (Maximum Anonymity)

Topic profiling is deactivated entirely. The node stores random, evenly distributed network fragments without recording query profiles, evicting strictly on FIFO (First-In, First-Out). Forensic inspection of your local disk reveals zero clues about your intellectual inquiries.

Ideal for zero-profile privacy

4. Protocol Boundaries & Wire Ceilings

Constraint Parameter	Hard Limit	Enforcement Rationale
Question Length	≤ 80 UTF-8 Characters	Atomic crystallization; prevents prompt bloat and guarantees exact cache hits.
Payload Ceiling	64 KB (65,536 Bytes)	Hard protocol ceiling preventing runaway log flooding or memory exhaustion.
Bandwidth Quota	1 MB/s In / 1 MB/s Out	Guarantees zero interference with operator's daily internet bandwidth.
Connection Cap	32 Total Peers / 2 per Peer	Resource conservation; operates comfortably on 15–25 MB RAM.
GossipSub Transport	128 KB Wire Frame	Framing envelope limit strictly enforced at the Rust libp2p perimeter.